+/-

This report is meant to help explore the results of the derfinder2 (Collado-Torres et al. 2013) package. While the report is rich, it is meant to just start the exploration of the results and exemplify some of the code used to do so. You will most likely need a more in-depth analysis for your specific data set.

Code setup

## For ggplot
tmp <- fullRegions
names(tmp) <- seq_len(length(tmp))
regions.df <- as.data.frame(tmp)
regions.df$width <- width(tmp)
rm(tmp)
nulls.df <- as.data.frame(fullNullSummary)

## Get chr lengths
if(hg19) {
    data(hg19Ideogram, package = "biovizBase")
    seqlengths(fullRegions) <- seqlengths(hg19Ideogram)[names(seqlengths(fullRegions))]
}

## Find which chrs are present in the data set
chrs <- levels(seqnames(fullRegions))
names(chrs) <- gsub("chr", "", chrs)

## Subset the fullCoverage data in case that a subset was used
colsubset <- optionsStats$colsubset
if(!is.null(fullCov) & !is.null(colsubset)) {
    fullCov <- lapply(fullCov, function(x) { x[, colsubset] })
}

## Get region coverage for the top regions
if(nBestRegions > 0) {
    regionCoverage <- getRegionCoverage(fullCov=fullCov, regions=fullRegions[seq_len(nBestRegions)], calculateMeans=FALSE, verbose=FALSE)
    save(regionCoverage, file=file.path(workingDir, "regionCoverage.Rdata"))
}

Quality checks

P-values

p1 <- ggplot(regions.df, aes(x=pvalues, colour=seqnames)) + geom_line(stat="density") + xlim(0, 1) + labs(title="Density of p-values") + xlab("p-values") + scale_colour_discrete(limits=chrs) + theme(legend.title=element_blank())
p1

## Compare the pvalues
summary(fullRegions$pvalues)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##  0.0006  0.0369  0.2460  0.3360  0.5780  1.0000

Q-values

summary(fullRegions$qvalues)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##  0.0107  0.0930  0.3150  0.3070  0.4950  0.6420

Cut	Count
0.0001	0
0.0010	0
0.0100	0
0.0250	2780
0.0500	3734
0.1000	4985
1.0000	19197

This table shows the number of candidate Differentially Expressed Regions (DERs) with q-value less than some commonly used cutoff values.

Region width

xrange <- range(log10(regions.df$width))
p2a <- ggplot(regions.df, aes(x=log10(width), colour=seqnames)) + geom_line(stat="density") + labs(title="Density of region lengths") + xlab("Region width (log10)") + scale_colour_discrete(limits=chrs) + xlim(xrange) + theme(legend.title=element_blank())
p2b <- ggplot(regions.df[idx.sig, ], aes(x=log10(width), colour=seqnames)) + geom_line(stat="density") + labs(title="Density of region lengths (significant only)") + xlab("Region width (log10)") + scale_colour_discrete(limits=chrs) + xlim(xrange) + theme(legend.title=element_blank())
grid.arrange(p2a, p2b)

This plot shows the density of the region lengths for all regions. The bottom panel is restricted to significant regions (q-value < 0.1)

Region Area

xrange <- range(log10(regions.df$area))
p3a <- ggplot(regions.df, aes(x=log10(area), colour=seqnames)) + geom_line(stat="density") + labs(title="Density of region areas") + xlab("Region area (log10)") + scale_colour_discrete(limits=chrs) + xlim(xrange) + theme(legend.title=element_blank())
p3b <- ggplot(regions.df[idx.sig, ], aes(x=log10(area), colour=seqnames)) + geom_line(stat="density") + labs(title="Density of region areas (significant only)") + xlab("Region area (log10)") + scale_colour_discrete(limits=chrs) + xlim(xrange) + theme(legend.title=element_blank())
grid.arrange(p3a, p3b)

This plot shows the density of the region areas for all regions. The bottom panel is restricted to significant regions (q-value < 0.1)

Null regions: width and area

p4 <- ggplot(nulls.df, aes(x=log10(width), colour=chr)) + geom_line(stat="density") + labs(title="Density of null region lengths") + xlab("Region width (log10)") + scale_colour_discrete(limits=chrs) + theme(legend.title=element_blank())
p5 <- ggplot(nulls.df, aes(x=log10(area), colour=chr)) + geom_line(stat="density") + labs(title="Density of null region areas") + xlab("Region area (log10)") + scale_colour_discrete(limits=chrs) + theme(legend.title=element_blank())
grid.arrange(p4, p5)

Mean coverage

xrange <- range(log2(regions.df$meanCoverage))
p6a <- ggplot(regions.df, aes(x=log2(meanCoverage), colour=seqnames)) + geom_line(stat="density") + labs(title="Density of region mean coverage") + xlab("Region mean coverage (log2)") + scale_colour_discrete(limits=chrs) + xlim(xrange) + theme(legend.title=element_blank())
p6b <- ggplot(regions.df[idx.sig, ], aes(x=log2(meanCoverage), colour=seqnames)) + geom_line(stat="density") + labs(title="Density of region mean coverage (significant only)") + xlab("Region mean coverage (log2)") + scale_colour_discrete(limits=chrs) + xlim(xrange) + theme(legend.title=element_blank())
grid.arrange(p6a, p6b)

This plot shows the density of the region mean coverage for all regions. The bottom panel is restricted to significant regions (q-value < 0.1)

Mean coverage vs fold change

The following plots are MA-style plots comparing each group vs the first one. The mean coverage is calculated using only two groups at a time and is weighted according to the number of samples on each group. Note that the mean coverage and fold change as calculated here do not taking into account the library sizes.

These plots are only shown when there are two or more groups. A total of 1 plot(s) were made.

for(j in grep("foldChange", colnames(values(fullRegions)))) {
    ## Identify the groups
    groups <- strsplit(gsub("foldChange", "", colnames(values(fullRegions))[j]), "vs")[[1]]
    
    ## Calculate the mean coverage only using the 2 groups in question
    j.mean <- which(colnames(values(fullRegions)) %in% paste0("mean", groups))
    groups.n <- sapply(groups, function(x) { sum(optionsStats$groupInfo == x) })
    ma.mean.mat <- as.matrix(values(fullRegions)[, j.mean])
    ## Weighted means
    ma.mean <- drop(ma.mean.mat %*% groups.n) / sum(groups.n)
    
    ma <- data.frame(mean=ma.mean, foldChange=values(fullRegions)[, j])
    ma2 <- ma[is.finite(ma$foldChange), ]
    fold.mean <- data.frame(foldMean=mean(ma2$foldChange, na.rm=TRUE))
    
    p.ma <- ggplot(ma, aes(x=log2(mean), y=foldChange)) + geom_point(size=1.5, alpha=1/5) + ylab("Fold Change (log2)\nRed dashed line at mean; blue line is loess fit") + xlab(paste("Mean coverage (log2) using only groups", groups[1], "and", groups[2])) + labs(title=paste("MA style plot:", groups[1], "vs ", groups[2])) + geom_hline(aes(yintercept=foldMean), data=fold.mean, colour="#990000", linetype="dashed") + geom_smooth(aes(y=foldChange, x=log2(mean)), data=subset(ma2, mean > 0), method=loess)
    print(p.ma)
}

Genomic overview

Q-values

plotOverview(regions=fullRegions, type="qval", base_size=30, areaRel=10, legend.position=c(0.97, 0.12))

This plot shows the genomic locations of the candidate regions found in the analysis. The significant regions (q-value less than 0.1) are highlighted and the area of the regions is shown on top of each chromosome. Note that the area is in a relative scale.

Annotation

plotOverview(regions=fullRegions, annotation=fullRegions, type="annotation", base_size=30, areaRel=10, legend.position=c(0.97, 0.12))

This genomic overview plot shows the annotation region type for the candidate regions. Note that the regions are shown only if the annotation information is available. Below is a table of the actual number of results per annotation region type.

Annotation (significant)

Region	Count
upstream	1894
promoter	94
overlaps 5'	8
inside	11710
overlaps 3'	5
close to 3'	27
downstream	1695
covers	0
	3764

plotOverview(regions=fullRegions[idx.sig], annotation=fullRegions[idx.sig], type="annotation", base_size=30, areaRel=10, legend.position=c(0.97, 0.12))

This genomic overview plot shows the annotation region type for the candidate regions that have a q-value less than 0.1. Note that the regions are shown only if the annotation information is available.

Best regions

Plots

Below are the plots for the top 100 candidate DERs as ranked by area. For each plot, annotation is shown if the candidate DER has a minimum overlap of 20 base pairs with annotation information (strand specific). If present, exons are collapsed and shown in blue. Introns are shown in light blue. The title of each plot is composed of the name of the nearest annotation element, the distance to it, and whether the region of the genome the DER falls into; all three pieces of information are based on bumphunter::annotateNearest().

The annotation depends on the Genomic State used. For details on which one was used for this report check the call to mergeResults in the reproducibility details.

if(nBestRegions > 0) {
    plotRegionCoverage(regions=fullRegions, regionCoverage=regionCoverage, groupInfo=optionsStats$groupInfo, nearestAnnotation=regions.df, annotatedRegions=fullAnnotatedRegions, whichRegions = seq_len(min(nBestRegions, length(fullRegions))), colors=NULL, scalefac = 32, ask = FALSE, verbose = TRUE) 
}

info <- do.call(rbind, lapply(fullAnnotatedRegions$countTable, function(x) { data.frame(table(x)) }))
colnames(info) <- c("Number of Overlapping States", "Frequency")
info$State <- gsub("\\..*", "", rownames(info))
rownames(info) <- NULL
print(xtable(info, align=rep("c", 4)), type="html", include.rownames=FALSE)

Region information

Number of Overlapping States	Frequency	State
0	17765	exon
1	1339	exon
2	93	exon
0	18224	intragenic
1	973	intragenic
0	18457	intron
1	708	intron
2	32	intron

Below is an interactive table with the top 2000 regions as ranked by area. This table was generated using rCharts (Vaidyanathan, 2013). Inf and -Inf are shown as 1e100 and -1e100 respectively.

topArea <- head(regions.df, 2000)
topArea <- data.frame("areaRank"=order(topArea$area, decreasing=TRUE), topArea)
## Clean up -Inf, Inf if present
## More details at https://github.com/ramnathv/rCharts/issues/259
replaceInf <- function(df, colsubset=seq_len(ncol(df))) {
    for(i in colsubset) {
        inf.idx <- !is.finite(df[, i])
        if(any(inf.idx)) {
            inf.sign <- sign(df[inf.idx, i])
            df[inf.idx, i] <- inf.sign * 1e100
        }
    }
    return(df)
}
topArea <- replaceInf(topArea, grep("foldChange", colnames(topArea)))

## Make the table
d <- data.table(topArea)
t1 <- dTable(d, sPaginationType=  'full_numbers', iDisplayLength=10, sScrollX='100%')
t1$print("regions", cdn=TRUE)

Best region clusters

Plots

## Select clusters by cluster area
df <- data.frame(area=fullRegions$area, clusterChr=paste0(as.integer(fullRegions$cluster), chr=as.character(seqnames(fullRegions))))
regionClustAreas <- tapply(df$area, df$clusterChr, sum)
bestArea <- sapply(names(head(sort(regionClustAreas, decreasing=TRUE), nBestClusters)), function(y) { which(df$clusterChr == y)[[1]]})

## Graphical setup: ideograms 
if(hg19 & is.null(p.ideos)) {
    ideos.set <- as.character(unique(seqnames(fullRegions[bestArea])))
    p.ideos <- lapply(ideos.set, function(xx) { 
        plotIdeogram(genome = "hg19", subchr=xx)
    })
    names(p.ideos) <- ideos.set
} else {
    stopifnot(!is.null(p.ideos))
}

## Graphical setup: transcription database
if(hg19 & is.null(txdb)) {
    txdb <- TxDb.Hsapiens.UCSC.hg19.knownGene
} else {
    stopifnot(!is.null(txdb))
}

## Graphical setup: main plotting function
regionClusterPlot <- function(idx, tUse="qval") {
    ## Chr specific selections
    chr <- as.character(seqnames(fullRegions[idx]))
    chrnum <- gsub("chr", "", chr)
    p.ideo <- p.ideos[[chr]]
    covInfo <- fullCov[[chrnum]]
    
    ## Make the plot
    p <- plotCluster(idx, regions=fullRegions, annotation=regions.df, coverageInfo=covInfo, groupInfo=optionsStats$groupInfo, titleUse=tUse, txdb=txdb, p.ideogram=p.ideo)
    print(p)
    rm(p.ideo, covInfo)
    
    return(invisible(TRUE)) 
}

Below are the best 20 region clusters ordered by cluster area (sum of the area of regions inside a cluster). The region with the highest area in the cluster is shown with a red bar.

## Genome plots
for(idx in bestArea) {
    regionClusterPlot(idx, ifelse(nullExist, "qval", "none"))
}

$plot of chunk bestClusters$

Permutations

Summary

## Get the permutation information
nSamples <- seq_len(length(optionsStats$groupInfo))
permuteInfo <- lapply(seeds, function(x) {
    set.seed(x)
    idx <- sample(nSamples)
    data.frame(optionsStats$groupInfo[idx])
})
permuteInfo <- cbind(data.frame(optionsStats$groupInfo), do.call(cbind, permuteInfo))
colnames(permuteInfo) <- c("original", paste0("perm", 1:10))
## The raw information
# permuteInfo

n <- names(table(permuteInfo[, 2]))
permuteDetail <- data.frame(matrix(NA, nrow=10*length(n), ncol = 2 + length(n)))
permuteDetail[, 1] <- rep(1:10, each=length(n))
permuteDetail[, 2] <- rep(n, 10)
colnames(permuteDetail) <- c("permutation", "group", as.character(n))
l <- 1
m <- 3:ncol(permuteDetail)
for(j in n) {
    k <- which(permuteInfo[, 1] == j)
    for(i in 2:11) {
        permuteDetail[l, m] <- table(permuteInfo[k, i])
        l <- l + 1
    }
}

## Print the summary
summary(permuteDetail[, m])

##       CEU            YRI   
##  Min.   : 5.0   Min.   :1  
##  1st Qu.: 7.0   1st Qu.:3  
##  Median :10.5   Median :5  
##  Mean   :10.5   Mean   :5  
##  3rd Qu.:14.0   3rd Qu.:7  
##  Max.   :16.0   Max.   :9

This table shows the summary per group of and can be used for fast detection of anomalies.

Note that in derfinder2 (Collado-Torres et al. 2013) the re-sampling of the samples is done without replacement. This is done in an effort to avoid singular model matrices. While the sample balance is the same across the permutations, what changes are the adjusted variables (including the column medians).

Interactive

The following table shows how the group labels were permuted. This can be useful to detect whether a permutation in particular had too many samples of a group labeled as another group, meaning that the resulting permuted group label resulted in pretty much a name change.

Bibliography

Reproducibility

General information

The F-statistic cutoff used was 0.00001 and type of cutoff used was theoretical. Furthermore, the maximum region (data) gap was set to 0 and the maximum cluster gap was set to 3000.

Details

This analysis was on each chromosome was performed with the following call to analyzeChr() (shown for one chromosome only):

## analyzeChr(chrnum = opt$chr, coverageInfo = data, testvars = group, 
##     adjustvars = adjustvars, cutoffFstat = 0.00001, cutoffType = "theoretical", 
##     nPermute = 10, seeds = 1:10, maxClusterGap = 3000, subject = "hg19", 
##     mc.cores = opt$mcores, verbose = opt$verbose)

## mergeResults(prefix = "run1-v0.0.24", genomicState = GenomicState.Hsapiens.UCSC.hg19.knownGene)

This report was generated in path /amber2/scratch/lcollado/derfinderExample/derAnalysis using the following call to generateReport():

## generateReport(prefix = "run1-v0.0.24", browse = FALSE, nBestClusters = 20, 
##     fullCov = fullCov, device = "CairoPNG")

## [1] "2013-09-25 23:21:21 EDT"

## Time difference of 24.65 mins

## R version 3.0.1 Patched (2013-05-16 r62754)
## Platform: x86_64-unknown-linux-gnu (64-bit)
## 
## locale:
##  [1] LC_CTYPE=en_US.iso885915       LC_NUMERIC=C                  
##  [3] LC_TIME=en_US.iso885915        LC_COLLATE=en_US.iso885915    
##  [5] LC_MONETARY=en_US.iso885915    LC_MESSAGES=en_US.iso885915   
##  [7] LC_PAPER=C                     LC_NAME=C                     
##  [9] LC_ADDRESS=C                   LC_TELEPHONE=C                
## [11] LC_MEASUREMENT=en_US.iso885915 LC_IDENTIFICATION=C           
## 
## attached base packages:
## [1] grid      parallel  methods   stats     graphics  grDevices utils    
## [8] datasets  base     
## 
## other attached packages:
##  [1] rtracklayer_1.20.2                     
##  [2] rCharts_0.3.53                         
##  [3] knitrBootstrap_0.8.0                   
##  [4] markdown_0.6.3                         
##  [5] RColorBrewer_1.0-5                     
##  [6] xtable_1.7-1                           
##  [7] knitcitations_0.4-7                    
##  [8] bibtex_0.3-6                           
##  [9] knitr_1.4.1                            
## [10] data.table_1.8.8                       
## [11] gridExtra_0.9.1                        
## [12] ggplot2_0.9.3.1                        
## [13] TxDb.Hsapiens.UCSC.hg19.knownGene_2.9.0
## [14] GenomicFeatures_1.12.1                 
## [15] AnnotationDbi_1.22.5                   
## [16] Biobase_2.20.0                         
## [17] GenomicRanges_1.12.3                   
## [18] IRanges_1.18.1                         
## [19] BiocGenerics_0.6.0                     
## [20] derfinder2_0.0.24                      
## [21] RcppArmadillo_0.3.820                  
## [22] Rcpp_0.10.4                            
## 
## loaded via a namespace (and not attached):
##  [1] biomaRt_2.16.0          Biostrings_2.28.0      
##  [3] biovizBase_1.8.0        bitops_1.0-5           
##  [5] BSgenome_1.28.0         Cairo_1.5-2            
##  [7] cluster_1.14.4          colorspace_1.2-2       
##  [9] DBI_0.2-7               dichromat_2.0-0        
## [11] digest_0.6.3            evaluate_0.4.7         
## [13] formatR_0.9             ggbio_1.8.5            
## [15] gtable_0.1.2            Hmisc_3.10-1.1         
## [17] httr_0.2                labeling_0.1           
## [19] lattice_0.20-15         MASS_7.3-26            
## [21] munsell_0.4             plyr_1.8               
## [23] proto_0.3-10            qvalue_1.34.0          
## [25] RCurl_1.95-4.1          reshape2_1.2.2         
## [27] RJSONIO_1.0-3           Rsamtools_1.12.3       
## [29] RSQLite_0.11.3          scales_0.2.3           
## [31] stats4_3.0.1            stringr_0.6.2          
## [33] tcltk_3.0.1             tools_3.0.1            
## [35] VariantAnnotation_1.6.5 whisker_0.3-2          
## [37] XML_3.96-1.1            yaml_2.1.7             
## [39] zlibbioc_1.6.0

derfinder2 basic results exploration

Introduction

Code setup

Quality checks

P-values

Q-values

Region width

Region Area

Null regions: width and area

Mean coverage

Mean coverage vs fold change

Genomic overview

Q-values

Annotation

Annotation (significant)

Best regions

Plots

Region information

Best region clusters

Plots

Permutations

Summary

Interactive

Bibliography

Reproducibility

General information

Details